Thread (35 messages) 35 messages, 10 authors, 2023-02-20

Re: [PATCH 1/1] ice: Change assigning method of the CPU affinity masks

From: Yury Norov <yury.norov@gmail.com>
Date: 2023-02-08 16:39:36
Also in: linux-crypto, linux-rdma, lkml

On Wed, Feb 08, 2023 at 04:39:05PM +0100, Pawel Chmielewski wrote:
quoted hunk ↗ jump to hunk
With the introduction of sched_numa_hop_mask() and
for_each_numa_hop_mask(), the affinity masks for queue vectors can be
conveniently set by preferring the CPUs that are closest to the NUMA node
of the parent PCI device.

Signed-off-by: Pawel Chmielewski <redacted>
---
 drivers/net/ethernet/intel/ice/ice_base.c | 17 ++++++++++++++---
 1 file changed, 14 insertions(+), 3 deletions(-)
diff --git a/drivers/net/ethernet/intel/ice/ice_base.c b/drivers/net/ethernet/intel/ice/ice_base.c
index e864634d66bc..fd3550d15c9e 100644
--- a/drivers/net/ethernet/intel/ice/ice_base.c
+++ b/drivers/net/ethernet/intel/ice/ice_base.c
@@ -122,8 +122,6 @@ static int ice_vsi_alloc_q_vector(struct ice_vsi *vsi, u16 v_idx)
 	if (vsi->type == ICE_VSI_VF)
 		goto out;
 	/* only set affinity_mask if the CPU is online */
-	if (cpu_online(v_idx))
-		cpumask_set_cpu(v_idx, &q_vector->affinity_mask);
 
 	/* This will not be called in the driver load path because the netdev
 	 * will not be created yet. All other cases with register the NAPI
@@ -659,8 +657,10 @@ int ice_vsi_wait_one_rx_ring(struct ice_vsi *vsi, bool ena, u16 rxq_idx)
  */
 int ice_vsi_alloc_q_vectors(struct ice_vsi *vsi)
 {
+	cpumask_t *aff_mask, *last_aff_mask = cpu_none_mask;
 	struct device *dev = ice_pf_to_dev(vsi->back);
-	u16 v_idx;
+	int numa_node = dev->numa_node;
+	u16 v_idx, cpu = 0;
 	int err;
 
 	if (vsi->q_vectors[0]) {
@@ -674,6 +674,17 @@ int ice_vsi_alloc_q_vectors(struct ice_vsi *vsi)
 			goto err_out;
 	}
 
+	v_idx = 0;
+	for_each_numa_hop_mask(aff_mask, numa_node) {
+		for_each_cpu_andnot(cpu, aff_mask, last_aff_mask)
+			if (v_idx < vsi->num_q_vectors) {
+				if (cpu_online(cpu))
+					cpumask_set_cpu(cpu, &vsi->q_vectors[v_idx]->affinity_mask);
+				v_idx++;
+			}
                        
                        else
                                goto out;
+		last_aff_mask = aff_mask;
+	}
+
        out:
 	return 0;
 
 err_out:
-- 
2.37.3
Would it make sense to increment v_idx only if matched CPU is online?
It will create a less sparse array of vectors...

Thanks,
Yury
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help